Achieving Fairness at No Utility Cost via Data Reweighing with Influence
Authors: Peizhao Li, Hongfu Liu
ICML 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results reveal that previous methods achieve fairness at a non-negligible cost of utility, while as a significant advantage, our approach can empirically release the tradeoff and obtain cost-free fairness for equal opportunity. We demonstrate the cost-free fairness through vanilla classifiers and standard training processes, compared to baseline methods on multiple real-world tabular datasets. |
| Researcher Affiliation | Academia | 1Brandeis University. Correspondence to: Peizhao Li <peizhaoli@brandeis.edu>. |
| Pseudocode | Yes | Algorithm 1 No Utility-Cost Fairness via Data Reweighing |
| Open Source Code | Yes | Code available at https://github.com/brandeis-machinelearning/influence-fairness. |
| Open Datasets | Yes | We use the following real-world tabular datasets for experiments (Dua & Graff, 2017). We provide statistics in Appendix B. Adult. The Adult dataset (Kohavi & Becker)... Compas. The Compas dataset (Julia Angwin & Kirchner, 2016)... Communities and Crime. The Communities and Crime dataset (Redmond & Baveja, 2002)... German Credit. The German Credit dataset (Hofmann)... |
| Dataset Splits | Yes | We divide all the datasets into training set (60%), validation set (20%), and test set (20%), except for the Adult dataset that has a pre-defined split on training/validation/test set. |
| Hardware Specification | No | The paper does not specify the hardware used for experiments (e.g., GPU/CPU models, memory, or specific computing cluster details). |
| Software Dependencies | Yes | Linear programs in Algorithm 1 are solved using Gurobi (Gurobi Optimization, LLC, 2021) under an academic license. |
| Experiment Setup | Yes | The specific parameters of base models and input data are used exactly the same across all baselines and our methods. Input data are standardized by removing the mean and scaling to unit variance. ... ℓ2 reg. is L2 regularization strength for the Logistic Regression model, obtained by a grid search over the validation set. ... We set 1e-3 as the L2 regularization for Neural Networks as default. ...Adult: Log Reg EOP: β = 0.5, γ = 0.2; NN EOP: β = 0.5, γ = 0.2; Log Reg DP: β = 0.8, γ = 0.3; NN DP: α = 0.02. |